The at-risk signal is not missing. It is sitting in systems you already run, often weeks before anyone fills out a withdrawal form. Student retention analytics fails at getting that signal to a person far more often than it fails at the math: the signal shows up, no single person owns it, and the term ends. Before you buy a model that predicts attrition, find out how late your current answer arrives and who it arrives to. On a campus the stall looks like this. It is Tuesday of week six. A student success advisor has 340 students on her caseload and a spreadsheet that was exported on Friday. One of those students stopped opening a 9am course in week three, went from four courses to three in week four, and filed two help desk tickets about a financial aid hold in week five. Three systems. Three offices. Nobody has seen all three facts in one place, and the advisor is not going to, because the report on her desk is a course-level pass rate. Every fact in that story existed before the student decided anything. None of them reached a person who could act while helping was still easy. What is student retention analytics actually measuring? Student retention analytics is the use of data a campus already collects to understand which students are at risk of leaving, and why. Most of it measures the outcome. Term-over-term persistence, cohort graduation rate, DFW rate by course, a year-over-year comparison for the board. That work matters. Someone has to keep score, and accreditation asks. But scorekeeping answers a question nobody can act on. It tells you how many students you lost after you have lost them. Retention is measured by cohort. Students leave one at a time, on a Tuesday. The operational question is smaller and much harder: what changed about this person’s behaviour in the last ten days, and who needs to know today. That question does not need a better statistical method. It needs three feeds you already own to arrive at the same desk. The signal is already in three systems you run Course engagement and attendance. This moves fastest. A student who logged in daily for two weeks and then not at all for five days has changed. So has one who misses the first low-stakes submission in a course, which in most terms is the earliest observable thing a campus records. It is noisy. Somebody gets the flu. That is fine, because a signal that moves this fast and misfires this often only calls for a small response: a text, not a meeting. Registration transactions. This is the most decisive of the three and the most ignored. Dropping from full-time to part-time, reversing a late add, a hold that will block registration for next term. Nobody drops a course casually. Each of those rows is a decision the student has already made, sitting in the student information system as a timestamp, filed as an administrative event instead of a warning. Support and service tickets. This is where a student says the problem out loud. Aid holds, a bursar balance, a housing problem, an account lockout, a parking appeal. A student who opens three tickets in ten days is telling you something the grade book will not say for six more weeks. Most campuses close those tickets against a response-time target and never read them as a group. None of these is a prediction. They are records of things that already happened, which is exactly why they are worth more than a risk score. A risk score is a guess someone can dispute. A dropped course is a fact. Why the early signal never reaches a person Two reasons, and neither is technical. The first is that the three systems belong to three offices, and each one reports to its own leadership instead of to the other two. The registrar’s numbers go to the registrar. The service desk reports on ticket volume and time to close. Course activity lives with the people who run the learning platform. Every office is doing its job correctly and no one of them can see the whole student. The second is that a report has no owner. A screen is a place to look, and looking is nobody’s job at 4pm in week six when there are 340 names on the list. If a signal does not arrive at a named person with a named action attached, it is not an early alert, it is one more screen, however accurate it is. That is also the honest limit on the whole idea. An early signal does not tell you a student is leaving. It tells you a student changed. The response has to be proportional: a text, a two-minute check-in, an offer to look at the hold together. Full intervention plans built off week-three attendance will burn your advisors out and annoy students who were only sick. How to measure your own early-warning window this week You do not need a vendor to find out where you stand. You need twenty names and an afternoon. Pull twenty students who withdrew or stopped out last term. Twenty is enough. You are working backwards through cases, not running a study. For each one, write down the date the exit became official. Go back through course activity, registration transactions and ticket history. Write the date of the first clearly observable change. Subtract. That gap, in days, is the early-warning window you already had. Add one more column, the hard one. On that earliest date, which system held the fact, and which human being could have seen it? Then write one sentence for the most common pattern you find, in this shape: when this happens, this person contacts the student within this many hours. If you cannot write that sentence, the signal has nowhere to go and better student retention analytics will not give it one. Expect two things. The window is probably longer than you think. The ownership is probably worse than you think, because on the earliest date the fact lived in a system the advisor has no login for. A thirty-day window that never reaches a person is not an analytics problem. It is a problem with who gets told. The student in week six is not a data point yet. She is a person with a hold she does not understand, a course she has quietly stopped opening, and three weeks left in which a five-minute conversation would still work. By the time she shows up in the persistence report, she is a number in a closed cohort, and the only thing left to do with her is count her. The gap between those two moments is real and it is measured in weeks. On most campuses it is currently being spent waiting for Friday’s export. This week: Take twenty students who left last term, find the first observable change in attendance, registration or tickets, and count the days between that date and the withdrawal.